Legal claims defining the scope of protection. Each claim is shown in both the original legal language and a plain English translation.
1. A non-transitory computer-readable medium comprising instructions for: providing a trained speech recognition neural network comprising one or more layers, each layer having one or more nodes; transcribing speech audio by the speech recognition neural network; during transcription, generating one or more feature representations from a subset of nodes, each feature representation corresponding to an internal state of the speech recognition neural network at a particular timestamp during transcription, wherein each of the feature representations comprises a set of node values obtained from the outputs of the nodes in the subset of nodes; storing the one or more feature representations; receiving a first set of classification values corresponding to an audio training portion of the speech audio; training a classification model on a first set of feature representations corresponding to the audio training portion of the speech audio and the first set of classification values, the first set of feature representations comprising a first subset of the feature representations generated during the speech audio transcription; and determining a second set of classification values in an unclassified audio portion of the speech audio by inputting a second set of feature representations corresponding to the unclassified audio portion of the speech audio into the trained classification model, the second set of feature representations comprising a second subset of the feature representations generated during the speech audio transcription.
This invention relates to speech recognition and classification systems. The problem addressed is the need for accurate and efficient classification of speech audio, particularly in scenarios where manual labeling is impractical or time-consuming. The solution involves a neural network-based approach that leverages internal states of a speech recognition model to improve classification performance. The system uses a trained speech recognition neural network with multiple layers, each containing nodes that process speech audio. During transcription, the network generates feature representations from subsets of nodes, capturing internal states at specific timestamps. These representations are stored and later used for classification tasks. The system receives labeled training data (classification values) for a portion of the speech audio and trains a classification model using the corresponding feature representations. The trained model is then applied to unclassified portions of the audio, using their feature representations to predict classification values. This approach enables efficient classification by reusing intermediate data from the speech recognition process, reducing the need for additional labeled data and improving accuracy. The method is particularly useful in applications requiring real-time or semi-supervised learning, such as voice assistants, transcription services, and automated content analysis.
2. The non-transitory computer-readable medium of claim 1 , wherein the one or more feature representations comprise quantized output values of the subset of nodes.
The invention relates to machine learning systems, specifically to techniques for processing feature representations in neural networks. The problem addressed is the computational inefficiency and memory overhead associated with storing and processing high-dimensional feature representations in neural networks, particularly when these representations are derived from intermediate layers of the network. The invention provides a method for optimizing neural network operations by quantizing the output values of a subset of nodes in a neural network layer. Quantization reduces the precision of these values, converting them into lower-bit representations (e.g., from 32-bit floating-point to 8-bit integers). This reduces memory usage and computational complexity while maintaining acceptable model accuracy. The quantized output values are then used as feature representations for subsequent processing steps, such as classification or regression tasks. The method involves selecting a subset of nodes from a neural network layer, processing their outputs to generate feature representations, and applying quantization to these representations. The quantization step may involve techniques such as uniform quantization, non-uniform quantization, or learned quantization to balance accuracy and efficiency. The quantized feature representations are then stored or transmitted for further use in the neural network pipeline. This approach is particularly useful in resource-constrained environments, such as edge devices or embedded systems, where computational efficiency and memory optimization are critical. By reducing the bit-width of feature representations, the method enables faster inference and lower power consumption without significant loss of model performance.
3. The non-transitory computer-readable medium of claim 1 , further comprising instructions for: applying one or more thresholds to the output of an activation function of each node in the subset of nodes and generating the feature representations from the resulting values.
The invention relates to machine learning systems, specifically to techniques for generating feature representations from neural network outputs. The problem addressed is improving the efficiency and accuracy of feature extraction in neural networks by selectively processing node outputs using thresholding techniques. The system involves a neural network with multiple layers of nodes, where each node applies an activation function to its input. A subset of nodes is selected based on their contribution to the final output or other criteria. For each node in this subset, the output of its activation function is processed by applying one or more thresholds. These thresholds filter or modify the node outputs, and the resulting values are then used to generate feature representations. The feature representations may be used for further processing, such as classification, regression, or other machine learning tasks. The thresholding step ensures that only significant or relevant node outputs contribute to the feature representations, reducing noise and computational overhead. The thresholds can be static or dynamically adjusted based on network performance or other factors. This approach enhances the interpretability and efficiency of neural network-based feature extraction.
4. The non-transitory computer-readable medium of claim 2 , wherein the quantized output values are thresholded output values of the subset of nodes.
The invention relates to neural network processing, specifically to techniques for optimizing neural network computations by quantizing and thresholding output values. Neural networks often require significant computational resources, particularly when deployed on edge devices with limited processing power. The invention addresses this challenge by reducing the computational load through quantization and thresholding of node outputs. The system processes a neural network by selecting a subset of nodes from the network. For each node in this subset, the system generates output values and then quantizes these values to reduce precision, which minimizes memory and computational requirements. The quantized values are further processed by applying a thresholding operation, converting them into binary or discrete values. This thresholding step simplifies subsequent computations, as binary or discrete values are easier to process than high-precision floating-point numbers. The thresholded output values are then used in further neural network operations, such as forward propagation or inference tasks. By reducing the precision of intermediate values and converting them into binary or discrete forms, the system significantly reduces the computational overhead while maintaining acceptable accuracy. This approach is particularly useful for real-time applications where low-latency processing is critical, such as in autonomous systems, mobile devices, or embedded systems. The invention improves efficiency without sacrificing the performance of the neural network.
5. The non-transitory computer-readable medium of claim 1 , wherein the subset of nodes includes nodes from a convolutional neural network layer and a fully connected neural network layer.
This invention relates to machine learning systems, specifically neural network architectures, and addresses the challenge of efficiently processing data through hybrid neural network structures. The invention involves a non-transitory computer-readable medium storing instructions for a neural network processing system that selects a subset of nodes from multiple neural network layers to optimize computational efficiency or accuracy. The subset includes nodes from both a convolutional neural network (CNN) layer and a fully connected neural network (FCN) layer, allowing the system to leverage the strengths of both layer types. The CNN layer processes spatial data hierarchically, while the FCN layer performs high-level feature extraction and classification. By combining nodes from these layers, the system can adapt to different tasks, such as image recognition or natural language processing, while reducing computational overhead. The instructions further enable dynamic selection of nodes based on input data characteristics, ensuring flexibility in handling diverse datasets. This approach improves performance by balancing accuracy and resource usage, making it suitable for applications requiring real-time processing or constrained environments. The invention enhances neural network efficiency by integrating heterogeneous layer types into a unified processing framework.
6. A non-transitory computer-readable medium comprising instructions for: providing a trained speech recognition neural network including a plurality of layers, each layer having a plurality of nodes; performing inference by the speech recognition neural network on input data comprising speech audio, wherein performing inference comprises transcribing the speech audio; while the speech recognition neural network is performing inference on the input data, generating one or more feature representations from a subset of nodes, the one or more feature representations corresponding to internal states of the speech recognition neural network at a plurality of timesteps, each feature representation corresponding to a list of values obtained from the output of the nodes in the subset of nodes; storing the one or more feature representations; receiving a first set of classification values for a training portion of the input data; training a learning model on a first set of feature representations corresponding to the training portion of the input data and the first set of classification values, each of the first set of feature representations comprising one of the lists of values generated during inference on the training portion of the input data; and determining a second set of classification values for an unclassified portion of the input data by inputting a second set of feature representations corresponding to the unclassified portion of the input data into the trained learning model, each of the second set of feature representations comprising one of the lists of values generated during inference on the unclassified portion of the input data.
This invention relates to speech recognition systems and the extraction of intermediate neural network features for downstream classification tasks. The problem addressed is the need to leverage internal states of a trained speech recognition neural network to improve classification accuracy for tasks beyond transcription, such as emotion detection or speaker identification. The system uses a pre-trained speech recognition neural network with multiple layers and nodes. During inference on speech audio input, the network transcribes the speech while simultaneously generating feature representations from a subset of its internal nodes. These features capture the network's internal states at multiple timesteps, with each representation being a list of values from the selected nodes. The extracted features are stored for further processing. A learning model is then trained using these features and corresponding classification labels for a labeled portion of the input data. The trained model is applied to unclassified portions of the input data by feeding the extracted features from those segments into the model, producing classification outputs. This approach enables the reuse of intermediate neural network representations for tasks that may not be directly addressed by the original speech recognition model. The method improves classification performance by leveraging the rich, task-agnostic features learned during speech recognition.
7. The non-transitory computer-readable medium of claim 6 , wherein the one or more feature representations comprise thresholded output values of the subset of nodes.
A system for processing data using a neural network includes a method for generating feature representations from a subset of nodes in the neural network. The method involves selecting a subset of nodes from the neural network, where the subset is determined based on a predefined criterion such as node importance or relevance to a specific task. The selected nodes produce output values, which are then processed to generate feature representations. These feature representations are derived by applying a thresholding operation to the output values of the subset of nodes, converting continuous or multi-level outputs into binary or discrete values. The thresholded output values serve as simplified or condensed representations of the original node outputs, which can be used for further analysis, classification, or decision-making tasks. This approach reduces computational complexity while preserving essential information from the neural network's internal state. The system is particularly useful in applications where interpretability, efficiency, or resource constraints are important, such as edge computing, real-time processing, or explainable AI systems. The method ensures that only the most relevant node outputs are retained, improving the system's performance and reliability.
8. The non-transitory computer-readable medium of claim 6 , further comprising instructions for: applying one or more thresholds to the output of an activation function of each node in the subset of nodes and generating the feature representations from the resulting values.
This invention relates to machine learning systems, specifically improving feature extraction in neural networks. The problem addressed is the inefficiency in generating meaningful feature representations from neural network outputs, which can lead to suboptimal performance in tasks like classification or regression. The solution involves a method for processing neural network outputs to enhance feature extraction. The system processes data through a neural network with multiple layers of nodes. A subset of nodes from these layers is selected based on predefined criteria, such as their contribution to the network's performance or their relevance to specific features. The outputs of these selected nodes are then refined by applying one or more thresholds to the results of an activation function. This step filters or normalizes the outputs, ensuring that only significant values are retained. The refined values are then used to generate feature representations, which are more discriminative and robust for downstream tasks. The thresholding step is critical as it reduces noise and irrelevant activations, improving the quality of the extracted features. This method can be applied to various neural network architectures, including convolutional or recurrent networks, to enhance their feature extraction capabilities. The resulting feature representations are better suited for tasks requiring high accuracy, such as image recognition, natural language processing, or anomaly detection.
9. The non-transitory computer-readable medium of claim 6 , wherein the one or more feature representations comprise quantized output values of the subset of nodes.
The invention relates to machine learning systems, specifically to techniques for processing neural network outputs. The problem addressed is the efficient representation and handling of feature data derived from neural networks, particularly in scenarios where computational efficiency or memory constraints are critical. The invention involves a method for storing and processing feature representations generated by a neural network, where these representations are derived from a subset of nodes within the network. The key innovation is the use of quantized output values from these nodes to form the feature representations. Quantization reduces the precision of the output values, which can significantly decrease storage requirements and computational overhead without substantially degrading the quality of the features. The quantized values are then used for further processing, such as classification or regression tasks, while maintaining the benefits of reduced resource usage. This approach is particularly useful in applications where neural networks are deployed on resource-constrained devices, such as mobile or embedded systems, where minimizing memory and processing demands is essential. The invention also includes techniques for selecting the subset of nodes whose outputs are quantized, ensuring that the most relevant features are retained while others are discarded to optimize performance. The overall system enables efficient neural network inference by leveraging quantized feature representations, balancing accuracy and computational efficiency.
10. The non-transitory computer-readable medium of claim 8 , wherein the subset of nodes includes nodes from a convolutional neural network layer and a fully connected neural network layer.
A system and method for optimizing neural network training involves selecting a subset of nodes from a neural network for gradient computation during backpropagation. The selection is based on a criterion that evaluates the contribution of each node to the network's output, reducing computational overhead while maintaining accuracy. The subset includes nodes from both convolutional neural network layers and fully connected neural network layers. By dynamically adjusting the subset of nodes considered during backpropagation, the system minimizes redundant calculations, improving training efficiency without sacrificing model performance. The approach is particularly useful in large-scale neural networks where full gradient computation is computationally expensive. The method involves analyzing node contributions, selecting the most influential nodes, and computing gradients only for those nodes, thereby optimizing resource usage during training. This technique can be applied to various neural network architectures, including those with convolutional and fully connected layers, to enhance training speed and scalability.
11. A non-transitory computer-readable medium comprising instructions for: providing a trained speech recognition neural network including a plurality of layers each having a plurality of nodes; transcribing speech audio by the speech recognition neural network; during transcription, generating one or more feature representations from a subset of nodes corresponding to an internal state of the speech recognition neural network at a plurality of timestamps, the feature representations comprising a list of values obtained from the nodes in the subset of nodes; storing the one or more feature representations; training a second neural network to generate a search feature based on a query input, the search feature comprising a plurality of values; receiving a query and inputting the query to the second neural network to generate the search feature; determining a relationship between the search feature and a plurality of the one or more feature representations; selecting a feature representation having a relationship with the search feature that most closely matches a predetermined criteria; and outputting an indication of a portion of the speech audio corresponding to the feature representation having the relationship with the search feature that most closely matches the predetermined criteria.
This invention relates to speech recognition and search systems, addressing the challenge of efficiently locating specific portions of transcribed speech audio based on user queries. The system uses a trained speech recognition neural network with multiple layers of nodes to transcribe speech audio. During transcription, it generates feature representations from a subset of nodes corresponding to the internal state of the network at various timestamps. These feature representations are stored as lists of values derived from the selected nodes. A second neural network is trained to generate search features from query inputs, where each search feature consists of multiple values. When a query is received, it is processed by the second neural network to produce a search feature. The system then evaluates the relationship between this search feature and the stored feature representations to identify the most relevant match based on predefined criteria. The output indicates the portion of the speech audio corresponding to the selected feature representation, enabling precise retrieval of relevant audio segments. This approach enhances search accuracy by leveraging internal neural network states rather than relying solely on traditional transcription-based search methods.
12. The non-transitory computer-readable medium of 11 , wherein the one or more feature representations comprise thresholded output values of the subset of nodes.
The invention relates to machine learning systems, specifically to techniques for processing feature representations in neural networks. The problem addressed is the efficient extraction and use of feature representations from neural network layers, particularly when dealing with high-dimensional data where only a subset of nodes may be relevant for a given task. The system involves a non-transitory computer-readable medium storing instructions that, when executed, cause a processor to generate feature representations from a neural network. These feature representations are derived from a subset of nodes within the network, where the output values of these nodes are thresholded to produce binary or discrete outputs. The thresholding process converts continuous activation values into simplified representations, reducing computational complexity while preserving meaningful information. The subset of nodes may be selected based on relevance to a specific task, such as classification or regression, allowing the system to focus on the most informative features. The instructions further enable the system to process these thresholded feature representations for downstream tasks, such as input to another neural network layer or a decision-making module. The thresholding step ensures that only significant activations are retained, improving efficiency and interpretability. This approach is particularly useful in applications where computational resources are limited or where explainability of the model's decisions is important. The system may be applied in various domains, including computer vision, natural language processing, and predictive analytics, where feature selection and dimensionality reduction are critical.
13. The non-transitory computer-readable medium of claim 11 , further comprising instructions for: applying one or more thresholds to an output of an activation function of each node in the subset of nodes and generating the set of features from the resulting values.
A system and method for processing data using a neural network involves extracting features from intermediate layers of the network. The neural network includes multiple layers of nodes, where each node applies an activation function to its input. The system identifies a subset of nodes from one or more intermediate layers of the network. For each node in this subset, the system applies one or more thresholds to the output of the activation function, generating a set of features from the resulting values. These features are then used for further processing, such as classification, regression, or other machine learning tasks. The thresholds may be predefined or dynamically adjusted based on the network's performance. This approach allows for the extraction of meaningful features from intermediate layers, improving the network's ability to capture complex patterns in the data. The method is particularly useful in applications where interpretability or feature extraction is important, such as in medical imaging, natural language processing, or autonomous systems. The system may be implemented as a software module integrated into a larger machine learning framework or as a standalone tool for feature extraction.
14. The non-transitory computer-readable medium of claim 11 , wherein the one or more feature representations comprise quantized output values of the subset of nodes.
The invention relates to machine learning systems, specifically to techniques for processing neural network outputs. The problem addressed is the efficient representation and storage of neural network feature data, particularly in scenarios where computational or memory resources are constrained. Traditional neural networks generate high-dimensional feature representations, which can be resource-intensive to store and process. The invention provides a solution by quantizing the output values of selected nodes in a neural network, reducing the data size while preserving essential information. The system involves a neural network with multiple layers of nodes, where each node generates an output value. A subset of these nodes is selected for quantization, meaning their output values are converted into a reduced-precision format (e.g., lower-bit integers). This quantization process compresses the feature representations, making them more efficient to store and transmit without significantly degrading model performance. The quantized values can then be used for downstream tasks such as classification, regression, or further processing in a machine learning pipeline. The invention is particularly useful in edge computing, embedded systems, or any application where minimizing memory usage and computational overhead is critical. By quantizing only a subset of nodes rather than the entire network, the system balances efficiency and accuracy, ensuring that the most important features retain high precision while less critical data is compressed. This approach enables real-time processing and deployment of neural networks in resource-limited environments.
15. The non-transitory computer-readable medium of claim 13 , wherein the search feature comprises a vector of values.
A system and method for improving search functionality in digital applications involves using a vector of values to enhance search accuracy and relevance. The invention addresses the problem of inefficient or inaccurate search results in digital databases, particularly when dealing with complex or unstructured data. Traditional keyword-based search methods often fail to capture the contextual or semantic relationships between terms, leading to poor search performance. The system includes a search feature that utilizes a vector of values to represent search queries and database entries. This vector-based approach allows for more nuanced comparisons between queries and stored data, improving the ability to retrieve relevant results. The vectors are generated by processing input data through a machine learning model or other computational techniques, converting textual or numerical information into a high-dimensional space where similarities can be quantified. By comparing the vectors of search queries against the vectors of stored entries, the system can identify the most relevant matches based on their proximity in the vector space. This method is particularly useful in applications requiring semantic search, such as natural language processing, recommendation systems, or knowledge management tools. The use of vectors enables the system to handle ambiguous or context-dependent queries more effectively than traditional keyword matching. The invention also includes mechanisms for updating and refining the vectors over time to adapt to changing data or user preferences, ensuring sustained accuracy and relevance.
16. A non-transitory computer-readable medium comprising instructions for: providing a trained speech recognition neural network, the speech recognition neural network including a plurality of layers each having a plurality of nodes; performing inference by the speech recognition neural network on input data, wherein the input data comprises speech audio and performing inference comprises transcribing the speech audio; generating, while performing inference on the input data, one or more feature representations from a subset of nodes, the one or more feature representations corresponding to an internal state of the speech recognition neural network at a given timestamp and comprising a set of internal values obtained from the node outputs of the subset of nodes; storing the one or more feature representations; training a second neural network to generate a search feature based on a query input, the search feature comprising a plurality of values; receiving a query and inputting the query into the second neural network to generate the search feature; determining a relationship between the search feature and each of the one or more feature representations; selecting a feature representation having a relationship with the search feature that most closely matches a predetermined criteria; and outputting an indication of a portion of the speech audio corresponding to the feature representation having the relationship with the search feature that most closely matches the predetermined criteria.
This invention relates to speech recognition and search systems, addressing the challenge of efficiently locating specific portions of transcribed speech audio based on user queries. The system uses a trained speech recognition neural network with multiple layers of nodes to transcribe input speech audio. During inference, the network generates feature representations from a subset of nodes, capturing the internal state of the network at specific timestamps. These feature representations are stored for later retrieval. A second neural network is trained to generate search features from query inputs. When a user submits a query, the second network processes it to produce a search feature. The system then compares this search feature against the stored feature representations to determine relationships, selecting the feature representation that best matches the query based on predefined criteria. The system outputs the corresponding portion of the original speech audio, enabling precise retrieval of relevant segments. This approach leverages neural network outputs to enhance search accuracy and efficiency, improving the usability of transcribed speech data. The system avoids traditional keyword-based search limitations by utilizing learned feature representations, allowing for more context-aware and semantically meaningful retrieval.
17. The non-transitory computer-readable medium of 16 , wherein the one or more feature representations comprise thresholded output values of the subset of nodes.
The invention relates to machine learning systems, specifically to techniques for processing feature representations in neural networks. The problem addressed is the efficient extraction and use of feature representations from neural network layers, particularly when dealing with large-scale models where computational efficiency is critical. Traditional methods may not adequately handle the selection and processing of feature representations, leading to suboptimal performance or excessive resource consumption. The invention describes a non-transitory computer-readable medium storing instructions that, when executed, cause a computing device to process feature representations derived from a neural network. The feature representations are generated from a subset of nodes within the neural network, where the subset is selected based on specific criteria. The key innovation involves using thresholded output values of the subset of nodes to form the feature representations. Thresholding ensures that only significant or meaningful outputs are retained, reducing noise and improving computational efficiency. The thresholded values are then used for further processing, such as classification, regression, or other downstream tasks. This approach allows for more efficient and accurate feature extraction, particularly in scenarios where the neural network has a large number of nodes or layers. The method can be applied to various types of neural networks, including convolutional neural networks (CNNs), recurrent neural networks (RNNs), and transformers, to enhance performance while minimizing computational overhead.
18. The non-transitory computer-readable medium of claim 16 , further comprising instructions for: applying one or more thresholds to the output of the activation function of each node in the subset of nodes and generating the set of features from the resulting values.
A system and method for processing data using a neural network involves extracting features from intermediate layers of the network. The neural network includes multiple layers of nodes, where each node applies an activation function to its input. The system selects a subset of nodes from one or more intermediate layers of the network. The activation function outputs of these selected nodes are processed by applying one or more thresholds to generate a set of features. These features are then used for further analysis or decision-making. The thresholds may be predefined or dynamically adjusted based on the network's performance. This approach allows for the extraction of meaningful features from intermediate layers, improving the network's ability to capture relevant patterns in the data. The method is particularly useful in applications where intermediate layer outputs need to be interpreted or used for downstream tasks, such as classification, regression, or anomaly detection. The system ensures that the extracted features are robust and discriminative by applying appropriate thresholds to the activation function outputs.
19. The non-transitory computer-readable medium of claim 16 , wherein the one or more feature representations comprise quantized output values of the subset of nodes.
The invention relates to machine learning systems, specifically to techniques for processing feature representations in neural networks. The problem addressed is the computational inefficiency and memory overhead associated with storing and processing high-dimensional feature representations in neural networks, particularly when these features are derived from intermediate layers of the network. The solution involves a non-transitory computer-readable medium storing instructions that, when executed, cause a processor to generate feature representations from a subset of nodes in a neural network. These feature representations are quantized, meaning their output values are converted into discrete, lower-precision values to reduce computational complexity and memory usage. The quantized feature representations are then used for further processing, such as classification or regression tasks, while maintaining the network's accuracy. The system includes a neural network with multiple layers, where a subset of nodes from one or more layers is selected to generate the feature representations. These nodes are processed to produce output values, which are then quantized into a reduced set of discrete values. The quantization step may involve techniques such as uniform quantization, where values are rounded to the nearest predefined level, or other methods that minimize information loss. By quantizing the feature representations, the system reduces the memory footprint and computational requirements of the neural network, making it more efficient for deployment in resource-constrained environments. The approach ensures that the quantized representations retain sufficient information for accurate downstream tasks, balancing performance and efficiency.
20. The non-transitory computer-readable medium of claim 18 , wherein the search feature comprises a vector of values.
A system and method for enhancing search functionality in a computing environment involves a search feature that utilizes a vector of values to improve search accuracy and relevance. The search feature processes input data by converting it into a vector representation, where each element of the vector corresponds to a specific attribute or characteristic of the data. This vector-based approach allows for more precise matching and ranking of search results by comparing the input vector against stored vectors in a database. The system may include a database storing multiple vectors, each associated with a data entry, and a processing module that performs vector comparisons to identify the most relevant matches. The search feature may also incorporate machine learning techniques to refine the vector representations over time, adapting to user preferences and improving search performance. This approach is particularly useful in applications requiring high-dimensional data analysis, such as natural language processing, image recognition, or recommendation systems, where traditional keyword-based searches are insufficient. The vector-based search feature enables more nuanced and context-aware search capabilities, enhancing user experience and efficiency in retrieving relevant information.
Unknown
November 24, 2020
Browse 5M+ US patents with plain-English claim translations and AI-generated analysis.